基于3D U-Net模型实现T_(2)WI图像中前列腺的区域分割及临床验证

Regional segmentation of prostate and clinical evaluation in T_(2)WI image based on 3D U-Net model

目的:探讨基于3D U-Net模型进行T_(2)WI图像中前列腺自动区域分割的可行性并进行临床验证。方法:回顾性收集我院2019年6月-2020年1月间的288例因临床怀疑前列腺癌而行多参数磁共振成像(Multiparameter magnetic resonance imaging,mpMRI)患者,用于3D U-Net分割模型的训练(数据集A)。所有患者按照8∶1∶1的比例随机分为训练集(n=231)、调优集(n=29)和测试集(n=28)。以医生手工标注的前列腺区域(外周带、移行带、中央带、前纤维基质带、尿道)为参考标准,定量评估测试集中模型的Dice相似系数(Dice similarity coefficient,DSC)、Hausdorff表面距离(Hausdorff surface distance,HSD)和体积差。另收集2020年2-5月间的经病理证实的30例前列腺癌患者,用于模型的临床验证(数据集B)。由一名低年资影像医生对检出的癌灶进行定位,一次使用AI结果,一次不使用AI结果。以影像专家的定位为参考标准,评估低年资医生两次阅片的定位准确率。结果:测试集中,模型对前列腺外周带、移行带、中央带、前纤维基质带和尿道的分割效果较好,其DSC及HSD分别为0.80、0.89、0.52、0.63、0.79和10.66 mm、8.88 mm、9.87 mm、17.86 mm、9.64 mm。模型预测与手工标注测量的体积一致性高,其差值基本都位于95%一致范围(limits of agreement,LoA)之内。与直接阅片相比,模型可提高低年资影像医生对癌灶的定位准确率(93.93%vs 100%)。结论:基于3D U-Net模型可实现T_(2)WI图像中前列腺区域自动分割,用于癌灶的定位。

Objective:To explore the feasibility of automatic regional segmentation of prostate in T_(2)WI images based on 3D U-Net model and clinical verification.Methods:288 patients with clinically suspected prostate cancer who underwent mpMRI in our hospital from June 2019 to January 2020 were retrospectively collected for 3D U-Net development(Dataset A).All patients were randomly divided into training set(n=231),validation set(n=29)and testing set(n=28)according to the ratio of 8∶1∶1.The manual annotations of prostate area(peripheral zone,transitional zone,central zone,anterior fibrous stromal zone,urethra)were regarded as reference standard.The Dice similarity coefficient(DSC),Hausdorff surface distance(HSD)and volume difference of the model in the testing set were quantitatively evaluated.Another 30 patients with pathologically confirmed prostate cancer between February 2020 and May 2020 were collected for clinical validation(Dataset B).A junior radiologist detected and localized the cancerous lesion using AI results once and not using AI results once.The localization accuracy of the junior radiologist was evaluated based on a radiology expert.Results:In the testing set,the 3D U-Net model showed good segmentation performance on the peripheral zone,transitional zone,central zone,anterior fibrous zone and urethra,with DSC and HSD of 0.80,0.89,0.52,0.63,0.79 and 10.66 mm,8.88 mm,9.87 mm,17.86 mm and 9.64 mm,respectively.The volume predicted by the model is consistent with the volume measured by manual annotation,and the difference is basically within the limits of agreement(LoA).The model improved the accuracy of tumor location of the junior radiologist from 93.93%to 100%.Conclusion:The automatic regional segmentation of prostate in T_(2)WI based on 3D U-Net model is feasible and can be used for cancer localization.